Hyperspectral image noise reduction and its effect on spectral unmixing

In hyperspectral images (HSI), many of the spectral bands have low noise levels (LN bands) while some have high noise levels (junk bands). If a noise reduction algorithm is globally applied to the whole dataset, it usually affects the LN bands adversely. Therefore, we consider different criteria for denoising LN and junk bands. First, we discriminate between LN and junk bands using the spectral correlation between adjacent bands. After that, the mirror-extended curvelet transform is applied to all spectral bands. Next, each LN band is denoised using a soft thresholding technique, while a local noise reduction method is used for the junk bands, where the curvelet coefficients of adjacent LN bands are used to recover the junk bands. This targeted approach is prone to reduce spectral distortions during denoising compared to global denoising methods. This is shown in simulations where the proposed method is compared to a wavelet and a 3-dimensional Wiener filtering denoising algorithm. The proposed method outperforms the global methods in terms of PSNR. To assess the effect of spectral distortion, spectral unmixing is applied to the denoised results and reconstruction errors are compared, again showing the benefit of the proposed approach.